Prediction of Mechanical Properties and Optimization of Friction Stir Welded 2195 Aluminum Alloy Based on BP Neural Network

Abstract

Friction stir welding (FSW) is regarded as an important joining process for the next generation of aerospace aluminum alloys. However, the performance of the FSW process often suffers from low precision and a long test cycle. In order to overcome these problems, a machine learning model based on a backpropagation neural network (BPNN) was developed to optimize the FSW of 2195 aluminum alloys. A four-dimensional mapping relationship between welding parameters and mechanical properties of joints was established through the analysis and mining of FSW data. The intelligent optimization of the welding process and the prediction of joint properties were realized. The weld formation characteristics at different welding parameters were analyzed to reveal the metallurgical mechanism behind the mapping relationship of the process-property obtained by the BPNN model. The results showed that the prediction accuracy of the method proposed could reach 92%. The welding parameters optimized by the BPNN model were 1810 rpm, 105 mm/min, and 3 kN for the rotational speed, welding speed, and welding pressure, respectively. Under these conditions, the tensile strength of the joint was found to be 415 MPa, which deviated from the experimental value by 3.71%.